Just as video killed the radio star, so too will AI demolish writers, journalists, and editors. Legions of wordsmiths — from Fiverr freelancers to The New York Times reporters — may soon find themselves out of work. However, they will be defeated not by competitors overseas, but by algorithms.
To understand the future of writing and how AI writers might look like, we first need to look at the types of jobs that are already on the chopping block.
Writing jobs aren’t easily automated…right?
Automation doesn’t touch all jobs equally and it can be easily seen in the US. A walk through the post-industrial heartland of Ohio yields a very different picture than a city like San Francisco, with its high concentration of intellectual capital.
As tempting as it may be to blame foreign competition, it appears that the great job-killer isn’t outsourcing, but rather automation. After all, American manufacturing is doing quite well: manufacturing grew by almost 2.2 percent per year — far faster than the overall US economy, which grew by 1.6 percent in 2016.
It seems manufacturing is booming, even if flesh-and-blood workers aren’t. But why are factory jobs so susceptible to robots? Why not programmers in San Francisco or a writer in New York City?
It comes down to tasks. Organizations as diverse as Oxford, McKinsey, and PwC, have concluded that jobs which are most easily automated share several important traits. They must have repetitive routines and high predictability (think assembly lines or stocking boxes at large warehouses). Such roles are straightforward, with little need for adaptation or lateral thinking.
Conversely, jobs with a high degree of unpredictability and a need for complex problem-solving are far less likely to be surrendered to the machines. One handy little tool from NPR, which predicts your profession’s chances of being automated, gives writers and authors a 3.8 percent chance of being edged out by computer programs. As the conventional wisdom goes, creativity isn’t easily replicated by machines.
Or is it?
The Turing test
For an AI writer to be effective, its work must pass the Turing test, in which a computer must trick humans into thinking that it too, is human.
This is particularly important for creative algorithms. Users don’t want to consume content created by a bot, as we believe that robots can’t effectively connect with us on an emotional level. We believe there is no formula for creativity: one cannot simply reduce a work like War and Peace to algorithms and binary inputs.
But the reality is that programmers can actually conjure creativity — and they’ve already done it. Back in 2011, an undergrad at Duke University modified an algorithm to dissect poems into smaller components (stanzas, lines, phrases) and then generate its own poems automatically. One was even accepted by Duke’s literary journal, The Archive. The AI writer thereby effectively passed the Turing test by passing off its own creation as a work by a human.
Of course there’s a world of difference between a poem of nine lines and a longform article in The New York Times (or some much more respected outlet, like TNW). However, it’s important to realize that this marks an important milestone; for years, people have speculated that creativity was beyond the reach of machines. Now that AIs have written poems, songs, and even short films — the writing is on the wall.
What would a robot writer look like?
Perhaps one reason that AI writers are hard to imagine is that the vast majority of them can’t perform at the same level as human writers, yet. For example, Facebook shut down its language-building AIs because they couldn’t use natural language effectively.
But to ignore AI because of a few public mishaps is dangerous, as the pieces of the puzzle are already in place. Not only have AI writers already passed the Turing test, but they can also rely on specialized algorithms like deep learning (which recently enabled an AI to defeat a human in the notoriously abstract game of Go) to hone their writing skills. Further, AIs can already process reams of data seamlessly, without the food and rest required by meat-based counterparts.
For instance, despite initial setbacks, IBM’s Watson has the ability to analyze thousands of reports and generate insights, even helping doctors fine tune diagnoses — and save lives in the process.
From here, it’s a small step to robot writers. In advertising, AI copywriters are highly versatile: they can draft hundreds of different ad campaigns, test and analyze the strengths of each different iteration, and utilizing deep learning, become better writers quickly. Most of all, AI won’t have to rest, get paid, or incur expenses like award shows.
To avoid going the way of the horse, it’s clear that we need a whole new paradigm of AI-human cooperation, not competition.
In fact, there’s already one in place at The Washington Post, where executives turned to AI writer Heliograf to help grow their web audience. Editors input keywords and templates into Heliograf on various events and outcomes. Heliograf then trawls the web for data and keyword matches; from that, it generates reports, or alerts reporters to double-check data anomalies for potential scoops.
Heliograf stories are straightforward reports on events like elections or Olympic competitions. They are not, however, in-depth analyses — a deliberate decision on the part of the Post. Rather than using a handful of well-researched longform stories to capture splintered, niche audiences, the Post uses Heliograf to create a flood of small, simple stories to attract pageviews.
As an augmented intelligence which works alongside humans, Heliograf is a more positive model of human-machine interaction. There’s still space for humans to research and write in-depth features, from coverage of maternal mortality in America to an undercover investigation of private prisons.
Even if machine learning enables AI to match human writing abilities and sift through reams of data, the human-interest angle (and interviewing) will prove harder to master. Journalists may do less straightforward reporting, and more higher-level analysis and investigation.
Still, pain might be unavoidable: later versions of Heliograf may trigger large layoffs, as newspapers cut costs by shedding local reporters and sports writers, even as they retain (or increase) their investigative staff. The manufacturing industry may offer some parallels: automation increased output, expensive human workers were fired, and the remaining jobs required increasingly advanced degrees or experience. It came down to numbers: a human welder costs $25 per hour (with benefits and vacation days), whereas robots cost only about $8 an hour after installation, maintenance, and operating expenses.
About the future of writing, one thing is absolutely clear: AI writers are already here. However, how extensive the damage will be — and how many layoffs and human suffering we can expect — remains unclear. It’s true that there’s a kernel of hope in the augmented intelligence model of interaction, but that doesn’t mean writers should rest easy. In a decade or two, writers may find themselves in the same predicament as factory workers today.